Tenstorrent has closed a $1.5 billion Series E at a $6.2 billion post-money valuation. Trace Capital and Innovation X led the round. The Santa Clara company, run by chip architect Jim Keller, roughly doubled its valuation from the previous round.
The investor list is the interesting part. Fidelity, T. Rowe Price, Morgan Stanley and ARK Invest sit next to trading firms Virtu Financial and XTX Markets, the stablecoin issuer Tether, and Sandisk, a memory maker. Agostinelli Group, Eclipse, Maverick Silicon, Protagonist, Troluce Capital, Cadence Ventures and Brett Barna also took part. That’s crossover money, the kind that usually turns up a round or two before a listing.
What the money is for
Tenstorrent named three uses. First, scaling deployment of its Galaxy servers. Second, serving open-weight models, both large language models and video generation. Third, its licensing business, where it sells the TT-Ascalon RISC-V CPU and Tensix AI cores to companies that design their own silicon. Samsung and LG are among the licensees.
Keller’s pitch is one architecture sold at every size. The same design goes into a single card, a workstation, a rack or a supercluster, or into another company’s chip as licensed IP. The software stack is open source, so developers can run frontier models locally and enterprises can keep them on premises.
Why it matters next to Nvidia
Nobody funds an AI chip startup to beat Nvidia at its own game this year. The bet is on the edges of the market. Some buyers want to own the hardware and the software outright. Others design their own silicon and need cores they can license. A third group wants models running off the cloud for cost or privacy reasons.
Tenstorrent covers all three with one product line. That’s the hedge. If the Galaxy server business grows slowly, the licensing business can still grow alongside every custom chip program that picks RISC-V. If licensing stalls, there’s hardware revenue.
Protagonist, an existing investor, said it was increasing its commitment after video generation results that beat industry benchmarks. That claim comes from an investor talking its own book. Still, video generation is compute-hungry and less locked to CUDA than training is, so it’s a sensible place for an alternative to show up first.
What to watch
Three numbers will say whether the valuation holds. How many Galaxy systems ship and to whom. How many licensees take Ascalon and Tensix into production silicon, not just evaluation. And whether open-weight model serving turns into recurring revenue or stays a showcase.
The round gives Tenstorrent room to find out. It doesn’t answer the question.
